|
| 1 | +# using Pkg |
| 2 | +# Pkg.activate(temp=true) |
| 3 | +# Pkg.add(url="https://github.com/christiangnrd/Metal.jl/", rev="MPSGraph") |
| 4 | +# Pkg.add(["GPUArrays", "Plots"]) |
| 5 | + |
| 6 | +# Uncomment if you want to compare with CPU |
| 7 | +# Pkg.add(["AppleAccelerate"]) |
| 8 | +# using AppleAccelerate |
| 9 | + |
| 10 | +using Metal, GPUArrays, LinearAlgebra, Printf |
| 11 | + |
| 12 | +using Plots |
| 13 | +using Plots.Measures |
| 14 | + |
| 15 | +n_gpu_cores = "??" |
| 16 | +# Comment this out if scary. Please mention number of cores in your comment when uploading the figure |
| 17 | +system_prof = read(`system_profiler SPDisplaysDataType`, String) |
| 18 | +n_gpu_cores = only(match(r"Total Number of Cores:\s*(\d+)", system_prof).captures) |
| 19 | + |
| 20 | +PLOT_TITLE = "Matmul peakflops for $(device().name) ($n_gpu_cores GPU cores)" |
| 21 | + |
| 22 | +function cpupeakflops(; n::Integer=4096, |
| 23 | + n_batch::Integer=1, |
| 24 | + inT::DataType=Float32, |
| 25 | + outT::DataType=inT, |
| 26 | + ntrials::Integer=4, |
| 27 | + verify=true) |
| 28 | + t = Base.zeros(Float64, ntrials) |
| 29 | + n_batch == 1 || @warn "n_batch > 1 not supported for `mul!`, running with n_batch=1" |
| 30 | + n_batch = 1 |
| 31 | + shape = (n, n) |
| 32 | + for i=1:ntrials |
| 33 | + c = zeros(outT, shape...) |
| 34 | + a = ones(inT, shape...) |
| 35 | + b = ones(inT, shape...) |
| 36 | + t[i] = @elapsed mul!(c, a, b) |
| 37 | + verify && @assert only(unique(Array(c))) == n |
| 38 | + end |
| 39 | + |
| 40 | + return n_batch*2*Float64(n)^3 / minimum(t) |
| 41 | +end |
| 42 | +function _peakflops(f, n, n_batch, inT, outT, ntrials; verify=true) |
| 43 | + t = Base.zeros(Float64, ntrials) |
| 44 | + shape = n_batch == 1 ? (n, n) : (n, n, n_batch) |
| 45 | + for i=1:ntrials |
| 46 | + c = mtl(zeros(outT, shape...)) |
| 47 | + a = mtl(ones(inT, shape...)) |
| 48 | + b = mtl(ones(inT, shape...)) |
| 49 | + t[i] = @elapsed Metal.@sync f(c, a, b) |
| 50 | + verify && @assert only(unique(Array(c))) == n |
| 51 | + end |
| 52 | + |
| 53 | + return n_batch*2*Float64(n)^3 / minimum(t) |
| 54 | +end |
| 55 | +function gpuarrpeakflops(; n::Integer=4096, |
| 56 | + n_batch::Integer=1, |
| 57 | + inT::DataType=Float32, |
| 58 | + outT::DataType=inT, |
| 59 | + ntrials::Integer=3, |
| 60 | + verify=true) |
| 61 | + n_batch == 1 || @warn "n_batch > 1 not supported for `GPUArrays.generic_matmatmul!`, running with n_batch=1" |
| 62 | + _peakflops(n, 1, inT, outT, ntrials; verify) do c, a, b |
| 63 | + GPUArrays.generic_matmatmul!(c, LinearAlgebra.wrap(a, 'N'), LinearAlgebra.wrap(b, 'N'), 1, 0) |
| 64 | + end |
| 65 | +end |
| 66 | +function mpspeakflops(; n::Integer=4096, |
| 67 | + n_batch::Integer=1, |
| 68 | + inT::DataType=Float32, |
| 69 | + outT::DataType=inT, |
| 70 | + ntrials::Integer=3, |
| 71 | + verify=true) |
| 72 | + _peakflops(MPS.matmul!, n, n_batch, inT, outT, ntrials; verify) |
| 73 | +end |
| 74 | +function graphpeakflops(; n::Integer=4096, |
| 75 | + n_batch::Integer=1, |
| 76 | + inT::DataType=Float32, |
| 77 | + outT::DataType=inT, |
| 78 | + ntrials::Integer=3, |
| 79 | + verify=true) |
| 80 | + _peakflops(MPSGraphs.graph_matmul!, n, n_batch, inT, outT, ntrials; verify) |
| 81 | +end |
| 82 | +function anepeakflops(; kwargs...) |
| 83 | + # VERY HACKY |
| 84 | + newDesc = MPSGraphs.MPSGraphCompilationDescriptor() |
| 85 | + # Use optimization level 0 to avoid operations being moved to the neural engine |
| 86 | + newDesc.optimizationLevel = MPSGraphs.MPSGraphOptimizationLevel1 |
| 87 | + |
| 88 | + oldDesc = MPSGraphs._default_exec_desc[].compilationDescriptor |
| 89 | + |
| 90 | + MPSGraphs._default_exec_desc[].compilationDescriptor = newDesc |
| 91 | + res = graphpeakflops(; kwargs...) |
| 92 | + MPSGraphs._default_exec_desc[].compilationDescriptor = oldDesc |
| 93 | + |
| 94 | + return res |
| 95 | +end |
| 96 | + |
| 97 | +function compare(Ns, Fs, inT, outT=inT; n_batch=1, ntrials) |
| 98 | + results = Dict() |
| 99 | + |
| 100 | + newFs = if (outT == Float16 || (outT == Float32 && inT == Float16)) |
| 101 | + Fs |
| 102 | + else |
| 103 | + filter(x -> !occursin("ANE", x[2]),Fs) |
| 104 | + end |
| 105 | + |
| 106 | + for (_, info_str) in newFs |
| 107 | + results[info_str] = Float64[] |
| 108 | + end |
| 109 | + |
| 110 | + prefixstr = "\33[2K\r($inT, $outT) " |
| 111 | + @time "$((inT, outT))" begin |
| 112 | + for n in Ns |
| 113 | + verify = (n < maxintfloat(outT) && (inT != Float16 || (n < maxintfloat(inT)))) |
| 114 | + n_str = "$n: " |
| 115 | + for (f, info_str) in newFs |
| 116 | + print(prefixstr, n_str, info_str) |
| 117 | + push!(results[info_str], f(; inT, outT, n, n_batch, ntrials, verify)) |
| 118 | + GC.gc() |
| 119 | + end |
| 120 | + end |
| 121 | + print("\33[2K\r") |
| 122 | + end |
| 123 | + return results |
| 124 | +end |
| 125 | + |
| 126 | +function main(; Ns=[50, 64, 100, 128, 250, 256, 500, 512, 1000, 1024, 2000, 2048, 4000, 4096, 6000, 6144, 8000, 8192],#, 10000], |
| 127 | + Fs=[ |
| 128 | + (mpspeakflops, "MPS"), |
| 129 | + (graphpeakflops, "MPSGraph"), |
| 130 | + (anepeakflops, "MPSGraph (ANE)"), |
| 131 | + # (gpuarrpeakflops, "GPUArrays"), |
| 132 | + # (cpupeakflops, "CPU (AppleAccelerate)"), # Uncomment to test CPU performance |
| 133 | + ], |
| 134 | + n_batch=1, |
| 135 | + ntrials=5, |
| 136 | + outpath="", |
| 137 | + outtype="svg", |
| 138 | + plt_title=PLOT_TITLE) |
| 139 | + Ts=[ |
| 140 | + (Int8, Float16), |
| 141 | + (Int8, Float32), |
| 142 | + (Int16, Float32), |
| 143 | + (Float16, Float16), |
| 144 | + (Float16, Float32), |
| 145 | + (Float32, Float32), |
| 146 | + ] |
| 147 | + |
| 148 | + res = Dict() |
| 149 | + |
| 150 | + ylim_upper = 9e12 |
| 151 | + |
| 152 | + for (inT, outT) in Ts |
| 153 | + tmpres = compare(Ns, Fs, inT, outT; n_batch, ntrials) |
| 154 | + |
| 155 | + plt = plot(xlabel="N, n_batch=$(n_batch)", legendtitle="($inT, $outT)") |
| 156 | + for (res, (_, info_str)) in zip(tmpres,Fs) |
| 157 | + flops = tmpres[info_str] |
| 158 | + peakf = @sprintf("%.3e", maximum(flops)) |
| 159 | + if maximum(flops) > ylim_upper |
| 160 | + ylim_upper = maximum(flops) * 1.02 |
| 161 | + end |
| 162 | + plot!(plt, Ns, tmpres[info_str]; linewidth=1.5, label="$(peakf) peak: $info_str") |
| 163 | + end |
| 164 | + res[(inT,outT)] = (plt=plt, results=tmpres) |
| 165 | + end |
| 166 | + |
| 167 | + finalplot = plot(res[Ts[1]].plt, res[Ts[2]].plt, res[Ts[3]].plt, res[Ts[4]].plt, res[Ts[5]].plt, res[Ts[6]].plt; layout=(2,3), |
| 168 | + ylim=(0,ylim_upper), |
| 169 | + plot_title=plt_title, |
| 170 | + tickfonthalign=:left, |
| 171 | + bottommargin=15pt, |
| 172 | + size=(2000,1200)) |
| 173 | + if !isnothing(outpath) |
| 174 | + savefig(plot(finalplot, dpi=500), joinpath(outpath, "bench_all_$(n_batch).$outtype")) |
| 175 | + end |
| 176 | + return res, finalplot |
| 177 | +end |
0 commit comments